libceph: stop allocating a new cipher on every crypto request

commit 7af3ea18 upstream.

This is useless and more importantly not allowed on the writeback path,
because crypto_alloc_skcipher() allocates memory with GFP_KERNEL, which
can recurse back into the filesystem:

    kworker/9:3     D ffff92303f318180     0 20732      2 0x00000080
    Workqueue: ceph-msgr ceph_con_workfn [libceph]
     ffff923035dd4480 ffff923038f8a0c0 0000000000000001 000000009eb27318
     ffff92269eb28000 ffff92269eb27338 ffff923036b145ac ffff923035dd4480
     00000000ffffffff ffff923036b145b0 ffffffff951eb4e1 ffff923036b145a8
    Call Trace:
     [<ffffffff951eb4e1>] ? schedule+0x31/0x80
     [<ffffffff951eb77a>] ? schedule_preempt_disabled+0xa/0x10
     [<ffffffff951ed1f4>] ? __mutex_lock_slowpath+0xb4/0x130
     [<ffffffff951ed28b>] ? mutex_lock+0x1b/0x30
     [<ffffffffc0a974b3>] ? xfs_reclaim_inodes_ag+0x233/0x2d0 [xfs]
     [<ffffffff94d92ba5>] ? move_active_pages_to_lru+0x125/0x270
     [<ffffffff94f2b985>] ? radix_tree_gang_lookup_tag+0xc5/0x1c0
     [<ffffffff94dad0f3>] ? __list_lru_walk_one.isra.3+0x33/0x120
     [<ffffffffc0a98331>] ? xfs_reclaim_inodes_nr+0x31/0x40 [xfs]
     [<ffffffff94e05bfe>] ? super_cache_scan+0x17e/0x190
     [<ffffffff94d919f3>] ? shrink_slab.part.38+0x1e3/0x3d0
     [<ffffffff94d9616a>] ? shrink_node+0x10a/0x320
     [<ffffffff94d96474>] ? do_try_to_free_pages+0xf4/0x350
     [<ffffffff94d967ba>] ? try_to_free_pages+0xea/0x1b0
     [<ffffffff94d863bd>] ? __alloc_pages_nodemask+0x61d/0xe60
     [<ffffffff94ddf42d>] ? cache_grow_begin+0x9d/0x560
     [<ffffffff94ddfb88>] ? fallback_alloc+0x148/0x1c0
     [<ffffffff94ed84e7>] ? __crypto_alloc_tfm+0x37/0x130
     [<ffffffff94de09db>] ? __kmalloc+0x1eb/0x580
     [<ffffffffc09fe2db>] ? crush_choose_firstn+0x3eb/0x470 [libceph]
     [<ffffffff94ed84e7>] ? __crypto_alloc_tfm+0x37/0x130
     [<ffffffff94ed9c19>] ? crypto_spawn_tfm+0x39/0x60
     [<ffffffffc08b30a3>] ? crypto_cbc_init_tfm+0x23/0x40 [cbc]
     [<ffffffff94ed857c>] ? __crypto_alloc_tfm+0xcc/0x130
     [<ffffffff94edcc23>] ? crypto_skcipher_init_tfm+0x113/0x180
     [<ffffffff94ed7cc3>] ? crypto_create_tfm+0x43/0xb0
     [<ffffffff94ed83b0>] ? crypto_larval_lookup+0x150/0x150
     [<ffffffff94ed7da2>] ? crypto_alloc_tfm+0x72/0x120
     [<ffffffffc0a01dd7>] ? ceph_aes_encrypt2+0x67/0x400 [libceph]
     [<ffffffffc09fd264>] ? ceph_pg_to_up_acting_osds+0x84/0x5b0 [libceph]
     [<ffffffff950d40a0>] ? release_sock+0x40/0x90
     [<ffffffff95139f94>] ? tcp_recvmsg+0x4b4/0xae0
     [<ffffffffc0a02714>] ? ceph_encrypt2+0x54/0xc0 [libceph]
     [<ffffffffc0a02b4d>] ? ceph_x_encrypt+0x5d/0x90 [libceph]
     [<ffffffffc0a02bdf>] ? calcu_signature+0x5f/0x90 [libceph]
     [<ffffffffc0a02ef5>] ? ceph_x_sign_message+0x35/0x50 [libceph]
     [<ffffffffc09e948c>] ? prepare_write_message_footer+0x5c/0xa0 [libceph]
     [<ffffffffc09ecd18>] ? ceph_con_workfn+0x2258/0x2dd0 [libceph]
     [<ffffffffc09e9903>] ? queue_con_delay+0x33/0xd0 [libceph]
     [<ffffffffc09f68ed>] ? __submit_request+0x20d/0x2f0 [libceph]
     [<ffffffffc09f6ef8>] ? ceph_osdc_start_request+0x28/0x30 [libceph]
     [<ffffffffc0b52603>] ? rbd_queue_workfn+0x2f3/0x350 [rbd]
     [<ffffffff94c94ec0>] ? process_one_work+0x160/0x410
     [<ffffffff94c951bd>] ? worker_thread+0x4d/0x480
     [<ffffffff94c95170>] ? process_one_work+0x410/0x410
     [<ffffffff94c9af8d>] ? kthread+0xcd/0xf0
     [<ffffffff951efb2f>] ? ret_from_fork+0x1f/0x40
     [<ffffffff94c9aec0>] ? kthread_create_on_node+0x190/0x190

Allocating the cipher along with the key fixes the issue - as long the
key doesn't change, a single cipher context can be used concurrently in
multiple requests.

We still can't take that GFP_KERNEL allocation though.  Both
ceph_crypto_key_clone() and ceph_crypto_key_decode() are called from
GFP_NOFS context, so resort to memalloc_noio_{save,restore}() here.
Reported-by: Lucas Stach <l.stach@pengutronix.de>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

net/ceph/crypto.c

@@ -13,14 +13,60 @@
 #include <linux/ceph/decode.h>
 #include "crypto.h"
 
+/*
+ * Set ->key and ->tfm.  The rest of the key should be filled in before
+ * this function is called.
+ */
+static int set_secret(struct ceph_crypto_key *key, void *buf)
+{
+	unsigned int noio_flag;
+	int ret;
+
+	key->key = NULL;
+	key->tfm = NULL;
+
+	switch (key->type) {
+	case CEPH_CRYPTO_NONE:
+		return 0; /* nothing to do */
+	case CEPH_CRYPTO_AES:
+		break;
+	default:
+		return -ENOTSUPP;
+	}
+
+	WARN_ON(!key->len);
+	key->key = kmemdup(buf, key->len, GFP_NOIO);
+	if (!key->key) {
+		ret = -ENOMEM;
+		goto fail;
+	}
+
+	/* crypto_alloc_skcipher() allocates with GFP_KERNEL */
+	noio_flag = memalloc_noio_save();
+	key->tfm = crypto_alloc_skcipher("cbc(aes)", 0, CRYPTO_ALG_ASYNC);
+	memalloc_noio_restore(noio_flag);
+	if (IS_ERR(key->tfm)) {
+		ret = PTR_ERR(key->tfm);
+		key->tfm = NULL;
+		goto fail;
+	}
+
+	ret = crypto_skcipher_setkey(key->tfm, key->key, key->len);
+	if (ret)
+		goto fail;
+
+	return 0;
+
+fail:
+	ceph_crypto_key_destroy(key);
+	return ret;
+}
+
 int ceph_crypto_key_clone(struct ceph_crypto_key *dst,
 			  const struct ceph_crypto_key *src)
 {
 	memcpy(dst, src, sizeof(struct ceph_crypto_key));
-	dst->key = kmemdup(src->key, src->len, GFP_NOFS);
-	if (!dst->key)
-		return -ENOMEM;
-	return 0;
+	return set_secret(dst, src->key);
 }
 
 int ceph_crypto_key_encode(struct ceph_crypto_key *key, void **p, void *end)
@@ -37,16 +83,16 @@ int ceph_crypto_key_encode(struct ceph_crypto_key *key, void **p, void *end)
 
 int ceph_crypto_key_decode(struct ceph_crypto_key *key, void **p, void *end)
 {
+	int ret;
+
 	ceph_decode_need(p, end, 2*sizeof(u16) + sizeof(key->created), bad);
 	key->type = ceph_decode_16(p);
 	ceph_decode_copy(p, &key->created, sizeof(key->created));
 	key->len = ceph_decode_16(p);
 	ceph_decode_need(p, end, key->len, bad);
-	key->key = kmalloc(key->len, GFP_NOFS);
-	if (!key->key)
-		return -ENOMEM;
-	ceph_decode_copy(p, key->key, key->len);
-	return 0;
+	ret = set_secret(key, *p);
+	*p += key->len;
+	return ret;
 
 bad:
 	dout("failed to decode crypto key\n");
@@ -85,14 +131,11 @@ void ceph_crypto_key_destroy(struct ceph_crypto_key *key)
 	if (key) {
 		kfree(key->key);
 		key->key = NULL;
+		crypto_free_skcipher(key->tfm);
+		key->tfm = NULL;
 	}
 }
 
-static struct crypto_skcipher *ceph_crypto_alloc_cipher(void)
-{
-	return crypto_alloc_skcipher("cbc(aes)", 0, CRYPTO_ALG_ASYNC);
-}
-
 static const u8 *aes_iv = (u8 *)CEPH_AES_IV;
 
 /*
@@ -168,8 +211,7 @@ static void teardown_sgtable(struct sg_table *sgt)
 static int ceph_aes_crypt(const struct ceph_crypto_key *key, bool encrypt,
 			  void *buf, int buf_len, int in_len, int *pout_len)
 {
-	struct crypto_skcipher *tfm = ceph_crypto_alloc_cipher();
-	SKCIPHER_REQUEST_ON_STACK(req, tfm);
+	SKCIPHER_REQUEST_ON_STACK(req, key->tfm);
 	struct sg_table sgt;
 	struct scatterlist prealloc_sg;
 	char iv[AES_BLOCK_SIZE] __aligned(8);
@@ -177,20 +219,15 @@ static int ceph_aes_crypt(const struct ceph_crypto_key *key, bool encrypt,
 	int crypt_len = encrypt ? in_len + pad_byte : in_len;
 	int ret;
 
-	if (IS_ERR(tfm))
-		return PTR_ERR(tfm);
-
 	WARN_ON(crypt_len > buf_len);
 	if (encrypt)
 		memset(buf + in_len, pad_byte, pad_byte);
 	ret = setup_sgtable(&sgt, &prealloc_sg, buf, crypt_len);
 	if (ret)
-		goto out_tfm;
+		return ret;
 
-	crypto_skcipher_setkey((void *)tfm, key->key, key->len);
 	memcpy(iv, aes_iv, AES_BLOCK_SIZE);
-
-	skcipher_request_set_tfm(req, tfm);
+	skcipher_request_set_tfm(req, key->tfm);
 	skcipher_request_set_callback(req, 0, NULL, NULL);
 	skcipher_request_set_crypt(req, sgt.sgl, sgt.sgl, crypt_len, iv);
 
@@ -232,8 +269,6 @@ static int ceph_aes_crypt(const struct ceph_crypto_key *key, bool encrypt,
 
 out_sgt:
 	teardown_sgtable(&sgt);
-out_tfm:
-	crypto_free_skcipher(tfm);
 	return ret;
 }
 

net/ceph/crypto.h

@@ -12,6 +12,7 @@ struct ceph_crypto_key {
 	struct ceph_timespec created;
 	int len;
 	void *key;
+	struct crypto_skcipher *tfm;
 };
 
 int ceph_crypto_key_clone(struct ceph_crypto_key *dst,